Advanced Pattern Recognition through Machine Learning for DAS Conveyor Condition Monitoring

The objectives of this project were to identify potential algorithms to allow machine learning to be used to render the system capable of making its own adjustments during commissioning and during operation.

A thorough literature review at the start revealed that the approaches that were assumed to be useful were not so, largely because techniques such as neural networks require a comprehensive accurate set of training data, also called “ground truth data”.  the machine learning techniques were altered to make use of unsupervised learning methods with clustering techniques being the primary method. Scatter plots were created in two and three dimensions for early visual inspection. A range of variables were chosen for the scatter plots such as spike frequencies, spike amplitude, spike location, natural bearing frequencies, fundamental roller rotation frequencies and their harmonics.

From the cluster analysis useful indicators were extracted that covered:

• Loss of laser power along the fibre itself;

• Spike and haystack height amplitude diminution;
• Specific frequencies that occurred more frequently;
• Roller slippage on the return belt increasing on steeper inclines;
• The spread of two different sized rollers along the belt.

The indicators obtained from the cluster analysis are now leading to useful algorithms such as:

• Automated laser power loss compensation along the belt;

• Automated checking of belt speed;
• Automated checking of the locations of different sized rollers;
• Improved identification of roller rotation harmonics, and, natural bearing frequencies;
• Improved separation of typical wear frequency patterns from the natural bearing acoustic frequencies.

Work is now continuing on converting the new algorithms into potential commercial program code for the next release of an improved commercial version of the system.